Stethoscope & Algorithms. Medical AI. Building https://xitter.cf/t.co/eShNnX3MIp.
pale blue dot
Joined April 2022
- Tweets6K
- Following2.2K
- Followers2.6K
- Likes24.1K
Pinned Tweet
This is now available. It's the most comprehensive collection of Nigerian treatment guidelines and covers 270 medical conditions. Can be used to build RAG pipelines or to validate your health AI application. You can access it here:
Chisom Rutherford retweeted
Just met with the super inspiring @rutherfordml
Kudos to you on @housejobng
There should be a benchmark that measures which model gives better advice. Maybe call it AdviceBench or something.
Wow
Muse Spark Open -- weights releasing today -- runs great on new Mac Studios thanks to work with @johnternus and Apple.
Chisom Rutherford retweeted
We just completed training Nerve-1, our first AI model built entirely from scratch.
Nerve-1 is a lightweight, ~34M-parameter model trained on over 23 billion tokens.
It’s designed for intelligent routing, deciding how much reasoning effort a task requires and directing it to the most suitable AI model.
It runs entirely on regular CPUs, making routing decisions in milliseconds. We estimate it could reduce our inference costs by at least 30%.
Rolling out soon to all businesses and users on @trydecideAI
This is how we've actually built housejob ng. It's very difficult to miss a diagnosis of have a blindspot when using it.
In building clinical AI, one thing you deal with is making sure the clinician considers all possible differentials without overwhelming them with caveats and exceptions.
Most people attach likelihood percentages to each diagnosis, with the least likely receiving the lowest percentage.
The problem with that is that LLMs are not good at calibration, so an LLM showing a likelihood of 5% may not really represent a statistical likelihood of 5% for the particular patient, given all the context, including the contexts the AI does not see or account for, like environment, demography, local epidemiology, etc.
Attaching context to each differential is helpful. But for a physician short on time, too much context can become another load they have to manage.
One method I use is to include the diagnoses that could be fatal at the top, even if the probability of them is low. Bringing them to the physician's conscious mind could also help them think more critically about whether they could be missing something important themselves.
Chisom Rutherford retweeted
Cardiology trainees: how do you hear about courses, conferences and regional teaching?
For most of us it’s WhatsApp groups and forwarded emails.
So I built Learn into Clinote. It puts it all in one place, and it’s free.
#CardioTwitter #MedEd
In building clinical AI, one thing you deal with is making sure the clinician considers all possible differentials without overwhelming them with caveats and exceptions.
Most people attach likelihood percentages to each diagnosis, with the least likely receiving the lowest percentage.
The problem with that is that LLMs are not good at calibration, so an LLM showing a likelihood of 5% may not really represent a statistical likelihood of 5% for the particular patient, given all the context, including the contexts the AI does not see or account for, like environment, demography, local epidemiology, etc.
Attaching context to each differential is helpful. But for a physician short on time, too much context can become another load they have to manage.
One method I use is to include the diagnoses that could be fatal at the top, even if the probability of them is low. Bringing them to the physician's conscious mind could also help them think more critically about whether they could be missing something important themselves.
Chisom Rutherford retweeted
Path and pharm students in Nigeria: help us test a new learning platform
We're looking for 6 medical students in their path and pharm year to try out PatandFam, a new online learning platform in development.
You'll join a video call for about an hour on Saturday 17 or Sunday 18 October, use the platform, and tell us what works and what doesn't.
You'll need a stable internet connection for the call. As a thank you, you'll receive ₦10000
Sign up by Wednesday 14 October: patandfam.com/testers
I noticed when most people are reporting their symptoms now, they usually say, "My ChatGPT told me X or Y" which means most patients are likely consulting ChatGPT before showing up to the hospital.
Maybe it's because my brain still associates hard work with the experience of struggling to understand/implement a task.
I'm getting more work done now than I ever have, but I keep feeling like I'm not doing any enough. Maybe it's because the cognitive labour needed to get things done is now almost down to zero.
Now I think of it...
imagine being that one person at anthropic who said "no we will not put our money in imagegen diffusion models" when the entire world including your biggest competitor pivoted to it.
to believe code itself would unlock image and video generation when code generation wasn't even that good yet is sheer conviction
Chisom Rutherford retweeted
No, they can't measure blood pressure accurately.
American Heart Statement: "currently available devices may be inappropriate for clinical use." and "using these devices to guide treatment could lead to inappropriate care and may pose safety risks to patients." @American_Heart
ahajournals.org/doi/10.1161/… open-access
gift link wsj.com/health/wellness/bloo…
